Application Of Neural Networks For Car License Recognition And Infrared Focal Plane Array Nonuniformity Correction

In recent years the artificial neural network(sANN) is widely used in many fields because of its strong ability of self-organized learning, fault-tolerance, robustness and non-linear processing advantages. Back propagation neural network (BPNN) is the most mature and widely used one in ANN solutions.In this paper some application problems when using BPNN and the influence on performance and general ability of the ANN are discussed firstly. The problems includes sample collection and pre-processing, topological structure determination, local minimum problem in training and learning process, momentum items addition, study factor determination, etc. Secondly all of these problems are researched and validated using matrix character recognition and function approximation experiment. Finally the BPNN is applied to car license recognition system and a car license recognition algorithm based on grey level normalization preprocessing ANN is given. Because in traditional car license recognition method the image is binary segmented before entering the ANN, some general binary segmentation methods have the problem in determining the limitation; when the scenery has uneven lighting, the binary segmentation will blur the character stroke and character information is lost during the process which brings difficult for following recognition work. Grey preprocessing algorithm can transform the character with different background and target grey level into character with similar background and target grey level. The recognition rate, reason of recognition failure and recognition adaptation are statistically analyzed and compared with PCA (primary component analyze) recognition method. The experiment shows that the algorithm is practical and can get better recognition performance. Also this paper applies the BPNN to infrared focal plane array (IRFPA) nonuniformity correction, proposes improved algorithm (momentum items addition, normalization self-tuning parameter and adaptive study factor) for fundamental ANN adaptive algorithm. The algorithm is tested by simulated infra-red image and the self-tuning performance is good. At last some lessons learned from application for the ANN is summarized.